.png&w=256&q=75)
1
1
Indonesia
2+ years of experience

CropWhisper is an end-to-end multimodal agentic AI system that helps smallholder farmers identify and treat crop diseases without needing an agronomist, a soil lab, or any agricultural training. The farmer provides photos of their affected crop, a short description of the problem, and their location. CropWhisper handles the rest. An 8-agent pipeline powered by Qwen3-VL and Qwen3 Reasoning — served via vLLM on AMD Instinct MI300X GPUs — works through the problem in stages. A Vision-Language agent produces a forensic, bias-free visual description of the crop. A Diagnosis agent cross-references that against the farmer's statement, real soil data from the ISRIC SoilGrids global database, and a RAG database of 190 confirmed disease cases. A Verification agent stress-tests the diagnosis. Finally, an Action Plan agent produces a prioritized report with immediate steps, cost estimates, a 7-day monitoring checklist, and escalation criteria. When confidence is low, the system asks for more evidence. A second 4-agent follow-up pipeline accepts new photos, re-analyzes the case, and returns a diff-style update explicitly tagging each change as CONTRADICT, MODIFY, or KEEP. This cascades until the diagnosis is confident. The production vision is a mobile app with real-time camera, voice I/O in the farmer's local language, GPS auto-detection, and offline-first triage — designed for 2G/3G and zero literacy requirements. Built on AMD Instinct MI300X, ROCm, vLLM, Qwen3-VL, Qwen3 Reasoning, LangGraph, and Supabase.
10 May 2026